UNIVERSITY OF MASSACHUSETTS Dept. of Electrical & Computer Engineering. Digital Computer Arithmetic ECE 666

Transcription

1 UNIVERSITY OF MASSACHUSETTS Dept. of Electrical & Computer Engineering Digital Computer Arithmetic ECE 666 Part 3 Sequential Algorithms for Multiplication and Division Israel Koren ECE666/Koren Part.3.1 Sequential Multiplication X, A - multiplier and multiplicand X=xn-1xn-2...x0 ; A=an-1an-2...a1a0 xn-1,an-1 - sign digits (sign-magnitude or complement methods) Sequential algorithm - n-1 steps Step j - multiplier bit xj examined; product xja (j) added to P - previously accumulated partial (0) product (P =0) -1 Multiplying by 2 - shift by one position to the right - alignment necessary since the weight of xj+1 is double that of xj ECE666/Koren Part.3.2 Page 1

2 Sequential Multiplication - Proof Repeated substitution =2 -(n-1) If both operands positive (xn-1=an-1=0) - The result is a product consisting of 2(n-1) bits for its magnitude ECE666/Koren Part.3.3 Number of product bits Maximum value of U - when A and X are maximal Last term positive for n 3, therefore 2n-2 bits required to represent the value - 2n-1 bits with the sign bit Signed-magnitude numbers - multiply two magnitudes and generate the sign separately (positive if both operands have the same sign and negative otherwise) ECE666/Koren Part.3.4 Page 2

3 Negative operands For two's and one's complement - distinguish between multiplication with a negative multiplicand A and multiplication with a negative multiplier X If only multiplicand is negative - no need to change the previous algorithm - only add some multiple of a negative number that is represented in either two's or one's complement ECE666/Koren Part.3.5 Multiplication - Example A - negative, two's complement, X - positive, 4 bits Product - 7 bits, including sign bit Registers are 4 bits long - a double-length register required for storing the final product Vertical line separates most from least significant half - each stored in a single-length register Bits in least significant half not used in the add operation ECE666/Koren Part.3.6 Page 3

4 Least significant half of product Only 3 bit positions are utilized - least significant bit position unused - not necessarily final arrangement The 3 bits can be stored in 3 rightmost positions Sign bit of second register can be set in two ways (1) Always set sign bit to 0, irrespective of sign of the product, since it is the least significant part of result (2) Set sign bit equal to sign bit of first register Another possible arrangement - Use all four bit positions in second register for the four least significant bits of the product Use the rightmost two bit positions in the first register Insert two copies of sign bit into remaining bit positions ECE666/Koren Part.3.7 Negative Multiplier - Two s Complement Each bit considered separately - sign bit (with negative weight) treated differently than other bits Two's complement numbers - ; If sign bit of multiplier is ignored - X A is the desired product - if xn-1=1 - a correction is necessary The multiplicand A is subtracted from the most significant half of U ECE666/Koren Part.3.8 Page 4

5 Negative Multiplier - Example Multiplier and multiplicand - negative numbers in two's complement In correction step, subtraction of multiplicand is performed by adding its two's complement ECE666/Koren Part.3.9 Negative Multiplier - One s Complement and If xn-1=1, start with P =A - this takes care of the second correction term xn-1 ulp A At the end of the process - subtract the first n-1 correction term xn-1 2 A (0) ECE666/Koren Part.3.10 Page 5

6 Negative Multiplier - Example Product of 5 and -3 - one's complement As in previous example - subtraction of first correction term - adding its one's complement Unlike previous example - one's complement has to be expanded to double size using the sign digit - a double-length binary adder is needed ECE666/Koren Part.3.11 Sequential Division Division - the most complex and time-consuming of the four basic arithmetic operations In general, result of division has two components Given a dividend X and a divisor D, generate a quotient Q and a remainder R such that X = Q D + R (with R < D) Assumption - X,D,Q,R - positive If a double-length product is available after a multiply and we wish to allow the use of this result in a subsequent divide, then X may occupy a double-length register, while all other operands stored in single-length registers ECE666/Koren Part.3.12 Page 6

7 Overflow & Divide by zero Q largest number stored in a single-length n-1 register (<2 for a register with n bits) n-1 1. X < 2 D - otherwise an overflow indication must be produced by arithmetic unit Condition can be satisfied by preshifting either X or D (or both) Preshifting is simple when operands are floatingpoint numbers 2. D 0 - otherwise - a divide by zero indication must be generated No corrective action can be taken when D=0 ECE666/Koren Part.3.13 Division Algorithm - Fractions Assumption - dividend, divisor, quotient, remainder are fractions - divide overflow condition is X<D Obtain Q=0.q1 qm (m=n-1) - sequence of subtractions and shifts Step i - remainder is compared to divisor D - if remainder larger - quotient bit qi=1, otherwise 0 ith step - ri = 2ri-1-qiD ; i=1,2,...,m ri is the new remainder and ri-1 is the previous remainder (r0=x) qi determined by comparing 2ri-1 to D - the most complicated operation in division process ECE666/Koren Part.3.14 Page 7

8 Division Algorithm - Proof The remainder in the last step is rm and repeated substitution of the basic expression yields Substituting r0=x and dividing both sides by 2 results in m -m hence rm2 = X - Q D as required True final remainder is R=rm2 -m ECE666/Koren Part.3.15 Division - Example 1 - Fractions X=( )2=1/2 D=(0.110)2=3/4 Dividend occupies double-length reg. X<D satisfied Generation of 2r0 - no overflow An extra bit position in the arithmetic unit needed Final result : Q=(0.101)2=5/8 -m -3 R=rm 2 =r32 =1/4 2 =1/32 Quotient and final remainder satisfy X=Q D + R = 5/8 3/4 + 1/32 = 16/32 = 1/2 Precise quotient is the infinite binary fraction 2/3= ECE666/Koren Part.3.16 Page 8

9 Division Algorithm - Integers Same procedure; Previous equation rewritten - (XF,DF,QF,RF -fractions) 2n-2 Dividing by 2 yields n-1 The condition X < 2 D becomes XF < DF X= =32; D=01102=6 n-1 Overflow condition X<2 D is tested by comparing the most significant half of X, 0100, to D, 0110 The results of the division are Q=01012 =5 and R=00102=2 In final step the true remainder R is generated - no -(n-1) need to further multiply it by 2 ECE666/Koren Part.3.17 Restoring Division Comparison - most difficult step in division If 2ri-1-D<0 - qi=0 - remainder restored to its previous value - restoring division Robertson diagram - shows that if ri-1<d, qi is selected so that ri<d Since r0=x<d - R<D m subtractions, m shift operations, an average of m/2 restore operations can be done by retaining a copy of the previous remainder - avoiding the time penalty ECE666/Koren Part.3.18 Page 9

10 Nonrestoring Division - Remainder Alternative - quotient bit correction and remainder restoration postponed to later steps Restoring method - if 2ri-1-D<0 - remainder is restored to 2ri-1 Then shifted and D again subtracted, obtaining 4ri-1-D - process repeated as long as remainder negative Nonrestoring - restore operation avoided Negative remainder 2ri-1-D<0 shifted, then corrected by adding D, obtaining 2(2ri-1-D)+D=4ri-1-D Same remainder obtained with restoring division ECE666/Koren Part.3.19 Nonrestoring Division - Quotient Correcting a wrong selection of quotient _ bit in step i - next bit, qi+1, can be negative - 1 If qi was incorrectly _ set to 1 - negative remainder - select qi+1=1 and add D to remainder Instead of qi qi+1=10 (too large) - qiqi+1=11=01 Further correction - if needed - in next steps _ Rule for qi : ECE666/Koren Part.3.20 Page 10

11 Nonrestoring Division - Diagram Simpler and faster than selection rule for restoring division - 2ri-1 compared to 0 instead of to D Same equation for remainder: ri = 2ri-1-qiD Divisor D subtracted if 2ri-1 >0, added if <0 ri-1 <D qi selected so ri <D qi 0 - at each step, either addition or subtraction is performed Not SD representation no redundancy in representation of quotient Exactly m add/subtract and shift operations ECE666/Koren Part.3.21 Nonrestoring Division - Example 1 X=( )2=1/2 D=(0.110)2=3/4 Final remainder - as before - Q=0.111=0.1012=5/8 Graphical representation Horizontal lines - add ±D Diagonal lines - multiply by 2 ECE666/Koren Part.3.22 Page 11

12 Nonrestoring Division - Advantage Important feature of nonrestoring division - easily extended to two's complement negative numbers Generalized selection rule for qi - Remainder changes signs during process - nothing special about a negative dividend X ECE666/Koren Part.3.23 Nonrestoring Division - Example 2 X=(0.100)2=1/2 D=(1.010)2=-3/ Final quotient - Q=0.111=0.1012= =-5/8 =1.011 in two s complement Final remainder = 1/32 - same sign as the dividend X ECE666/Koren Part.3.24 Page 12

13 Nonrestoring Division - sign of remainder Sign of final remainder - same as dividend Example - dividing 5 by 3 - Q=1, R=2, not Q=2, R=-1 (although R <D) If sign of final remainder different from that of dividend - correction needed - results - from quotient digits being restricted to 1,1 Last digit can not be 0 - an even" quotient can not be generated ECE666/Koren Part.3.25 Nonrestoring Division - Example 3 X=0.1012= 5/8 D=0.1102= 3/4 Final remainder negative - dividend positive Correct final remainder by adding D to r =0.100 Correct quotient - Qcorrected = Q - ulp Q= Qcorrected=0.1102=3/4 ECE666/Koren Part.3.26 Page 13

14 Nonrestoring Division - Cont. If final remainder and dividend have opposite signs - correction needed If dividend and divisor have same sign - remainder rm corrected by adding D and quotient corrected by subtracting ulp If dividend and divisor have opposite signs - subtract D from rm and correct quotient by adding ulp Another consequence of the fact that 0 is not an allowed digit in non-restoring division - need for correction if a 0 remainder is generated in an intermediate step ECE666/Koren Part.3.27 Nonrestoring Division - Example 4 X=1.1012=-3/8 D=0.1102=3/4 Correct result of division - Q=-1/2; R=0 Although final remainder and dividend have same sign - correction needed due to a zero intermediate remainder This must be detected and corrected - r3(corrected) = r3+d= = Correcting the quotient Q=0.111=0.101 by subtracting ulp : Q(corrected) = =-1/2 ECE666/Koren Part.3.28 Page 14

15 Generating a Two's Complement Quotient in Nonrestoring Division Converting from using 1,1 to two's complement Previous algorithms require all digits of quotient before conversion starts - increasing execution time Preferable - conversion on the fly - serially from most to least significant digit as they become available Quotient digit assumes two values - single bit - sufficient for representation - 0 and 1 assigned to 1 and 1 Resulting binary number - 0.p1...pm ( pi=1/2(qi+1) ) - ECE666/Koren Part.3.29 Conversion Algorithm Step 1: Shift number one bit position to left Step 2: Complement most significant bit Step 3: Shift a 1 into least significant position Result - (1-p1).p2p3...pm1 - has same numerical value as original quotient Q Proof: Value of above sequence in two's complement - Substituting pi=1/2(qi+1) - Last term = m ECE666/Koren Part.3.30 Page 15

16 Conversion Algorithm - Example Algorithm can be executed in a bit-serial - fashion - Example - X=1.101 ; D= Instead of generating the quotient bits generate the bits (1-0).101=1.101 After correction step - Q-ulp= correct representation of -1/2 in two's complement Exercise - The same on-the-fly conversion algorithm can be derived from the general SD to two's complement conversion algorithm presented before ECE666/Koren Part.3.31 Square Root Extraction - Restoring The conventional completing the square method for square root extraction is conceptually similar to restoring division X - the radicand - a positive fraction ; Q=(0.q1 q2...qm) - its square root The bits of Q generated in m steps - one per step - partially developed root at step i (Qm=Q) ; ri - remainder in step i Calculation of next remainder - Square root extraction can be viewed as division with a changing divisor - ECE666/Koren Part.3.32 Page 16

17 Square Root Extraction - Cont. First step - remainder=radicand X ; Q0=0 Performed calculation - To determine qi in the restoring scheme - calculate a tentative remainder q1.q2 qi-101 = 2Qi simple to calculate If tentative remainder positive - its value is stored in ri and qi=1 Otherwise - ri=2ri-1 and qi=0 -i ECE666/Koren Part.3.33 Proof of Algorithm Repeated substitution in the expression for rm - m Dividing by 2 results in the expected relation -m with rm2 as the final remainder ECE666/Koren Part.3.34 Page 17

18 Example - Square root (Restoring) X= =11/16=176/256 Q= =13/ Final remainder=2 r4=7/256=x-q =( )/256 ECE666/Koren Part.3.35 Different Algorithm - Nonrestoring Second algorithm - similar to nonrestoring division Example - X= =25/64 Square root - - Q=0.111 =0.1012=5/8 Converting the digits of Q to two's complement representation - similarly to nonrestoring division Faster algorithms for square root extraction exist ECE666/Koren Part.3.36 Page 18